JIT: Faster comparison against Vector128<>.Zero

[EgorBo]

A good first issue for those who are interested in SIMD and ARM64. Should improve #63285

bool IsZero(Vector128<byte> v) => v == Vector128<byte>.Zero;

Currently is suboptimal on ARM64:

while I'd expect it to emit something like this:

so if one of the args of == is _Zero it needs to use umaxv, it should be done here:

runtime/src/coreclr/jit/lowerarmarch.cpp

Line 864 in 386f871

comp->gtNewSimdHWIntrinsicNode(simdType, cmp, NI_AdvSimd_Arm64_MinAcross, CORINFO_TYPE_UBYTE, simdSize);

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

bool IsZero(Vector128<byte> v) => v == Vector128<byte>.Zero;

Currently is suboptimal on ARM64:

while I'd expect it to emit something like this:

so if one of the args of == is _Zero it needs to use umaxv, it should be done here:

runtime/src/coreclr/jit/lowerarmarch.cpp

Line 864 in 386f871

comp->gtNewSimdHWIntrinsicNode(simdType, cmp, NI_AdvSimd_Arm64_MinAcross, CORINFO_TYPE_UBYTE, simdSize);

Author:	EgorBo
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[robertDurst]

👀 looking, interested

[robertDurst]

@EgorBo what tool is that?

[echesakov]

The code is slightly better after #62933

; Assembly listing for method Program:<<Main>$>g__IsZero|0_0(System.Runtime.Intrinsics.Vector128`1[Byte]):bool
; Emitting BLENDED_CODE for generic ARM64 CPU - MacOS
; optimized code
; fp based frame
; partially interruptible
; No PGO data
; Final local variable assignments
;
;  V00 arg0         [V00,T00] (  3,  3   )  simd16  ->   d0         HFA(simd16)  single-def
;# V01 OutArgs      [V01    ] (  1,  1   )  lclBlk ( 0) [sp+00H]   "OutgoingArgSpace"
;
; Lcl frame size = 0

G_M25750_IG01:              ;; offset=0000H
        00000000          stp     fp, lr, [sp,#-16]!
        00000000          mov     fp, sp
						;; bbWeight=1    PerfScore 1.50
G_M25750_IG02:              ;; offset=0008H
        00000000          cmeq    v16.16b, v0.16b, #0
        00000000          uminv   b16, v16.16b
        00000000          umov    w0, v16.b[0]
        00000000          cmp     w0, #0
        00000000          cset    x0, ne
						;; bbWeight=1    PerfScore 6.00
G_M25750_IG03:              ;; offset=001CH
        00000000          ldp     fp, lr, [sp],#16
        00000000          ret     lr
						;; bbWeight=1    PerfScore 2.00

; Total bytes of code 36, prolog size 8, PerfScore 13.10, instruction count 9, allocated bytes for code 36 (MethodHash=80c69b69) for method Program:<<Main>$>g__IsZero|0_0(System.Runtime.Intrinsics.Vector128`1[Byte]):bool
; ============================================================

cc @TIHan